Research On Semi-supervised Classification Algorithm Based On Differential Evolution And Extreme Learning Machine

The combination of evolutionary algorithm and analytical method is a hot topic in the field of machine learning in recent years.Semi-supervised classification algorithm based on Extreme learning machine(ELM)is widely used in indoor location estimation and other fields.In this paper,we focus on combining the semi-supervised classification method based on ELM with differential evolution algorithm.The research on this issue has just started and few related works have been reported.The existing methods do not have enough high classification accuracy and can not meet the demands of market decision.To address the above problems,we propose a semi-supervised classification method based on differential evolution(DE)and ELM algorithm(DE-ELM-SSC).Compared with the latest modified ELM algorithm based on cooperative training and differential evolution(Tri-DE-ELM)method,DE-ELM-SSC according to the training data sets selects the differential evolution strategy.Its main steps are as follows.(1)An optimal strategy for the target data set is selected in three typical differential evolution strategies according to the Root mean square error(RMSE).(2)The optimal evolutionary strategy is applied by the DE algorithm.(3)We optimize the network parameters of ELM based on the DE algorithm using the optimal strategy.(4)The improved ELMs in the previous step are used as base classifiers to construct a semi-supervised classification prediction model by Tri-training.Although DE-ELM-SSC can improve the classification accuracy,it has the problem of fixed scaling factor.Hence,we choose a common scaling factor method without loss of generality,then adopt an exponential method to improve the adaptive function to get optimized DE-ELM-SSC~+algorithm,and can further improve the classification accuracy of the algorithm.Extensive experimental results on UCI data sets show that the DE-ELM-SSC~+algorithm takes same or less training time in most experiment data sets and outperforms the baseline methods with higher classification accuracy because it can choose proper strategies according to different datasets,and adaptively adjust the scaling factor.